Many excellent things were formulated in the 70s and didn't catch on because supporting technologies weren't practical yet. It was an excellent time to be a fresh PhD
The roots of deep learning are from around the same era, if not a bit earlier - it's just that we didn't have the computing power to do much with it so DL kind of languished until ~15 years ago with the advent of capable GPUs. Is ART important? Can it do what DL does now? Hard to say. A lot of money and effort has been poured into DL to get it this far. If ART had gotten the attention/investment instead would it be where DL is now?
Have they though? I think I've run into realizations that deep learning is too expensive or not yet scalable enough to be useful in some applications, but like other commenters on this thread. It can feel like a lot of the criticisms are grounded in some sort of feeling about what true intelligence should look like and an idea that number crunching can't possible be it
It depends on what you mean by machine learning, but in practice stuff like video game enemy AI generally avoids deep learning for many reasons behind cost or scale. Time, tuning, complexity, and consistency for example are major issues in the real world.
Sure the SCII AI eventually became very strong but it’s also somewhat immersion breaking to see all the odd stuff it ends up doing. And that’s with a well tuned game, in something under active development you end up with a host of other issues.
Those are good points. Worth noting that video game AIs also often cheat. The goal for a video game ai is less to play the game well then to convince the player that they played the game well, but I think there definitely is something about believability or maybe human likeness that weirdly sometimes deep systems get very badly wrong
Applying deep learning to computer game AI is just really hard. Not even a data center worth of compute and a big team of world leading scientists could create an AI able to beat the best players at games like starcraft or dota. It isn't reasonable to spend many times more more money making an AI for a game than you spent making the entire game, and the AI can't even run on any normal computer so you can't even use it at the end.
Yes though critically if you did want to actually write an AI capable of beating good players deep learning would probably be important. It has successfully been used to win at starcraft and dota now right?
Wait, weren't both Starcraft and Dota getting very good AI results a few years ago by deepmind's alphastar and openAI respectively, IIRC they had some competitions with world champions? And I believe for the purpose of generalization and research they intentionally made the problem harder by using the visual render/representation instead of obvious technical simplifications of just taking the game state as any normal game bot would do.
I mean, playing well is obviously not a goal for making the AI for a single-player mode or a bot for a multi-player game; the goal there is to play in an entertaining way, and the optimal system probably is one that is simple, easy to test, and is beatable in a manner that's fun for the players. Probably the only niche (very narrow!) where you would actually want near-optimal AI is for making a "sparring partner" for training e-sports players.
they got good results, but only in highly controlled, short duration testing. they're near the deep blue point, but still far from the era where humans are convincingly and resoundingly beaten.
A big aspect is that for everyone working on those competition systems, making an actually good game AI was not even a goal - the only purpose of these experiments is as a proof of concept (perhaps as PR) and a testbed for techniques which would then apply to other practical tasks simulating human decisions; and the research attention seems to have moved from popular strategy games to more specific simulated environments, which are better at being experimental testbeds i.e. faster/cheaper to simulate.
People working on AI systems that are going to be actually used in games shipped to gamers is a separate community with (IMHO) almost no overlap.
I can't speak for ART one way or the other as a technology, but this is a good time to remember that there's no particular reason to believe that a good model for training a machine learning system will map analogously to how human brains actually function.
For all we know, there are multiple ways to converge to a trainable input recognition system. And if ART is bringing comprehensibility to the table, it may look even less like what actual brains do than modern deep learning algorithms; nothing about what I know regarding how brains operate suggests to me that they are comprehensible as anything other than "a pattern that works."
Deep learning is very good at certain problems, much better (in some sense, particularly the ability to generalize with less fine tuning and feature engineering) than methods that came before it. It had undeniably opened up tons of new research applications, but has of course fallen short of the hype around it. I'm still very bullish on the post-hype potential of deep learning for solving real problems, it just requires a much clearer understanding of what it does and doesn't do, and adapting solutions to work within it's constraints.
All that to say, it's easy to find flaws with it, but it has been game changing. I'm naturally skeptical about someone who has just published a book on something hes been working on for 50 years that he believes is "better" than the current (admittedly fad) technology. I'd like to see some real examples about the specific kinds of problems his tech solves, and how it solves them better than other methods. If he's saying it's better than deep learning on things deep learning is good at (e.g. can it beat a neural network model at classifying imagenet?) then I call BS absent a compelling demo. If it's better at some other thing, let's hear what that is.
I think the progress was the result of competitions that kept everyone blindfolded to test datasets. No one could exaggerate the performance of their methods, allowing deep learning to emerge as a winning solution. I think deep learning has fallen short of the hype because researchers are exaggerating their results, and what is needed to propel progress is a renewed focus (with funding) for competitions.
Exactly. If it's really better than deep learning, he can publish a library implementing it. Plenty of us are hungry for high performance, stable, explainable ML and the place for ART's approach would quickly be recognized.
Off Topic : I am interested in learning DL, not superficially (like most of online courses), but a more theoretical foundation, one that enables me to think and envisage new architectures, new techniques or combinations. Can you help me with any self study stuff that will enable me to learn as such.
> but a more theoretical foundation, one that enables me to think and envisage new architectures, new techniques or combinations
Advances in DL have been mostly empirical not theoretical from what I have seen. It's less about being able to "envisage" and more about empirical testing.
Just my opinion, but look into the "Overparameterized machine learning" community[0]. If you are interested in the theory, this gets at what is fundamentally different about deep learning.
I may be reading too deeply into your question and you don't actually mean that theoretical. For something lighter, I'd suggest studying some of the common model architectures in the area you're interested in (like you can look at the pytorch lightning or torchvision or detectron2 models or etc source code) and understanding why things are the way they are.
[0] https://topml.rice.edu/ (this is an annual conference, can look at last years talks and references therein, search for topml, etc)
I’m struck by the juxtaposition of these cogent, reasonable assessments of new tech (ML, blockchain) I see a consensus on and looking for jobs on LinkedIn. Sometimes all I see are jobs for new companies based on ML/blockchain.
We have these tools that solve specific problems in specific contexts yet all these companies are on the hype marketing train. Are these investors not doing due diligence or am I missing something?
Ten+ years into the DL revolution and we are still getting shallow hit piece articles like this from IEEE.
Look, everyone will jump on the ART bandwagon if it ever actually beats deep learning. Machine Learning is a very empirical field! But all the people throwing rocks at Deep Learning have failed to propose anything that actually works better.
> Ten+ years into the DL revolution and we are still getting shallow hit piece articles like this from IEEE.
I think the latest IEEE articles have been a little intellectually lightweight, perhaps we're just simply not their audience. Still, in the style of IEEE papers, I would like to see references for the claims made.
> But all the people throwing rocks at Deep Learning have failed to propose anything that actually works better
I was thinking something similar. DL for sure has problems, but I am yet to see something better. ART doesn't just need to be better, it needs to be better in _at least_ one of: accuracy, training time, recall time, amount of training data, quality of training data.
Unfortunately Amazon reviews, at least in my experience, mean very little any more. There are just too many marketing firms that specialize in messing with them.
Anecdotal example: a few months ago I bought an electric shaver that was very inexpensive and had thousands of 5-star reviews. It failed two months later, and when I went to leave the negative review, I found that the product no longer existed, but another identical-looking product with a different name from a different(?) company was the best-reviewed cheap electric shaver.
The curve doesn't look legit. Percentages for each star level. I have the impression Amazon could predict potentially manipulated reviews with reasonable accuracy by analyzing the curve.
Edit: not claiming or implying anything about the author, book or the reviews. Just pointing the curve doesn't look legit to me.
"Stephen Grossberg explains why his ART model is better"
Please don't explain me why your model is better.
_show me_ your model is better with results.
I see a lot of: "this model is explainable and has no forgetting modes". But I cannot care about an explainable model that does not work.
There are many classic explainable models, how do they compare to art in real life applications?
"The problem with today’s AI, he says, is that it tries to imitate the results of brain processing instead of probing the mechanisms that give rise to the results."
Today's AI works really well for many applications (computer vision has leaped in quality). But of course I did not expect perfection from ImageNet trained models. I also did not expect gpt3 to be agi. But are they good (especially when compared with what we had not long ago)? Hell yeah
Also, the advancements in protein folding are superb.
I understand the author is not dismissing the current advancement. But if you claim your model is better you ought to compare it with the state of the art
The latest paper there is from 2009 which is before the popularity rise of deep learning, before the creation of key methods of modern deep learning, and before the creation of most key datasets on which we would nowadays evaluate and compare the advantages and drawbacks of different methods.
Literally nothing in that repository makes an argument that this method is better than deep learning in any aspect whatsoever, they don't even attempt such a comparison.
At least, there can be found some on the theory and methods.
Regarding a direct comparison, it may be difficult to do such a comparison by just running some code on data sets, which the one group of algorithms has been heavily optimized for with excessive funding, while the other has been not. Even more so, as such a review would have a difficult time to asses any of the critical features, like chances of memory loss. Comparing technologies isn't as easy as running them side by side.
I appreciate that you're saying something that you think is only common sense,
but I suspect that is most likely because 99% of the machine learning papers
you've seen do nothing else but claim a new SOTA result on some set of
benchmarks. Yet just because that's what (almost) everyone is doing doesn't mean
it's good research. Quite the contrary, that's the reason why machine learning
research has become a quagmire and progress has stalled.
Here's Geoff Hinton on benchmarks (again; I cite that all the time):
GH: One big challenge the community faces is that if you want to get a paper
published in machine learning now it's got to have a table in it, with all these
different data sets across the top, and all these different methods along the
side, and your method has to look like the best one. If it doesn’t look like
that, it’s hard to get published. I don't think that's encouraging people to
think about radically new ideas.
Now if you send in a paper that has a radically new idea, there's no chance in
hell it will get accepted, because it's going to get some junior reviewer who
doesn't understand it. Or it’s going to get a senior reviewer who's trying to
review too many papers and doesn't understand it first time round and assumes it
must be nonsense. Anything that makes the brain hurt is not going to get
accepted. And I think that's really bad.
What we should be going for, particularly in the basic science conferences, is
radically new ideas. Because we know a radically new idea in the long run is
going to be much more influential than a tiny improvement. That's I think the
main downside of the fact that we've got this inversion now, where you've got a
few senior guys and a gazillion young guys.
P.S. And, please leave that aggressive attitude out of research discussions
("bullshit walks"?). Science is not the constant comparison of dick measurements.
That's perhaps the point of sports, but anything that stifles debate and
inhibits the development of new ideas is anti-science.
Just last year there was a breakthrough in decades old problem of protein-folding. The year before that we figured out that large language models have astonishing few shot learning abilities. CLIP introduced strong 0-shot capabilities to image recognition. Progress in image generation from descriptions has been nothing but astonishing with Dall-E and GLIDE from OpenAI.
Not to mention breakthroughs in game theoretical problems like Counterfactual Regret Minimization and more recent application of Deep Learning to make CFR practical in actual games. Including continuous progress in the AlphaGo linage to AlphaZero, MuZero, and last year EfficientZero and Player of Games.
With the exception of counterfactual regret minimisation, which has nothing to
do with deep learning, all those are applications of existing approaches that
have become possible because of the increased expenditure of resources-
processing power and data. There have been no theoretical advances, no
algorithmic advances to speak of, no new knowledge learned in the last 20 years
of reserach in deep learning. All that has been achieved is new SOTA on old
benchmarks [1].
To make an analogy it's as if we could sort ever larger lists because we keep
sorting them with ever bigger computers... but the only sorting algorithm we
know is bubblesort.
You bring up large language models. Language models are about as old as
computational linguistics [6]. Large language models are... large. There's
nothing that's new about them. "Attention" is an anthropomorphically misnamed
engineering tweak of the kind that was rightly lambasted by Drew McDermot fourty
years ago [7]. Self-play and reinforcement Learning are now more than 60 years
old [8].
As to "few shot" learning, that is said for a language model previously trained
on a copy of the entire web! In what sense is that "few shot"? It is truly
shocking to hear researchers make such obviously false claims with a straight
face. And that they are rewarded with publication of their work is nothing less
than a scandal.
Deep learning research is going nowhere. It is stuck in a rut. It is dead. It
has given up the spirit. It is an ex-research field. This research field is
a goner. It is dead, jim. And I'm sure it will be decades before people start to
say it out loud (the current generation of reserachers will leave the dirty job
to their students), but that doesn't change a thing.
______________
[1] Except in identifying new weaknesses of deep learning systems, as for
example Overinterpretation in the paper I cite above, adversarial examples [2],
shortcut learning [3], "the elephant in the room" [4], being "right for the
wrong reasons" [5] etc etc etc etc.
Which of course are very welcome. The only way to address the limitations of
current approaches is to become aware of them and try to understand them.
Unfortunately, it is only a tiny, tiny minority in the field who does that,
while the majority is happy to spout nonsense like "[a billion plus] few shot"
learning etc.
[8] Arhtur Samuel's checkers player beat a human champion in 1961. Donald
Michie build MENACE, a reinforcement learning algorithm implemented in
matchboxes in 1960:
>> From my perspective, I don't care for theoretical advances. I care about results
and if scale provides results, I'm happy.
Like I say you can sort ever larger lists with ever bigger computers and
bubblesort. But you can go a lot further if you use the brain and come up with a
better sorting algorithm.
But I think you would be really surprised to find out how much humans have
achieved by using their brains rather than ever bigger computers. Here's a small
example that happened to be on my mind recently:
That is an example of the difference between expensive toys like Dall-E and
world-changing scientific work. Though I appreciate that the difference can
sometimes appear somewhat mudddled and a machine that can generate a "hamster
dragon" may seem as an important scientific breakthrough. Depends on what you're
more interested in: hamster dragons, or understanding how the world works. Oh
well.
As to this:
>> It's also interesting coming from someone researching Inductive Logic
Programming that provides no results at all.
That's an obvious -and odious- attempt at trolling me, but unlike yours, my HN
profile is linked to my real-world identity and so I am forced to respond with
politenes and to represent my field with dignity. Therefore, I will point out a
famous achievement of ILP with which you are sadly unfamiliar and that I expect
you to immediately discount on the grounds that you didn't know about it, so it
can't have been all that important. I'm linking to an article in the popular
press that will be more accessible. It doesn't go into depth over the machine
learning approach used but it's ILP; I link to the scholarly article immediately
after:
Robot Makes Scientific Discovery All by Itself
For the first time, a robotic system has made a novel scientific discovery with
virtually no human intellectual input. Scientists designed “Adam” to carry out
the entire scientific process on its own: formulating hypotheses, designing and
running experiments, analyzing data, and deciding which experiments to run
next.
>> That is an example of the difference between expensive toys like Dall-E and world-changing scientific work.
There isn't really that much difference. DeepMind took some 3 years to max out CASP protein-folding benchmark running from 1994 on a problem known since ~1960. Protein folding is exactly the type of world-changing scientific work. But it is just a tip of the iceberg. DL is being successfully applied to PDE solving or electron density predictions and more.
The psychedelic hamster dragon is an example of the fact that DL is capable of 0-shot generalization. It is a significant scientific observation in itself.
And with regard to your example. We are discussing whether deep learning has staled. I give you mostly examples from last year like EfficientZero (30 Oct) ~500x improvement in sample efficiency over 2013 DQN, some literally not older than a month: GLIDE (20 Dec), Player of Games (6 Dec).
You are giving me example from 2004 (2009?).
I'm not trying to troll you. I just think you are biased in you assessment of significant results and what it means to be stalled.
>> Protein folding is exactly the type of world-changing scientific work.
Nope. Not at all. That's a misunderstanding of the goal of science which is to
explain how the world works. AlphaFold, like all neural network models is a
predictive model but has no explanatory power. It can predict the structure of
proteins from sequences but it cannot explain why, or how, proteins fold.
Scientists can still not explain how proteins fold, certainly not by interacting
with AlphaFold.
Why is that a problem? You said you care about results. I linked to the
wikipedia article about the Copernican Revolution. I recommend you read the
wikipedia article on epicycles [1]. To summarise, for hundreds of years, from
Hipparchus of Nicaea, through Claudius Ptolemy and all the way up to Copernicus,
astronomers used systems of epicycles (circles-upon-circles) to explain the
apparent movements of the planets. Epicycle-based models matched observations
very well and predicted the movement of the planets very accuratey because it is
always possible to fit a smooth curve with arbitrary precision given a
sufficient number of epicycles. However, even Copernicus' model that had the sun
in the center of the universe, unlike many earlier models, could not explain
why the planets moved the way they did. A true explanation only became
available when Isaac Newton developed his laws of universal gravitation.
Newton's theory (itself later superseded by Einstein's theory of general
relativity) provided a true explanation of observable phenomena and a way to
calculate better models and obtain more accurate predictions than were ever
possible before. That is the power of explaining and understanding the world and
simply modelling an incomprehensible set of observations, which is all that
neural networks can do, doesn't even come close.
>> I'm not trying to troll you. I just think you are biased in you assessment of
significant results and what it means to be stalled.
ILP is a tiny field with maybe a few dozen people that consistently publish on
the subject. By comparison, there are many tens of thousands of papers published
on deep learning every year, the deep learning conferences receive many
thousands of submissions and research is funded with many millions of dollars by
governments, militaries, and the largest of large technology corporatations, the
Googles and Facebooks etc, and training state-of-the-art models requires many
terrabytes of data, many petaflops of computing power, to the extent that
training such models is now only feasible for the aforementioned large
corporations. Normally, I'd complain that you are comparing apples to oranges.
How can a few people with meager resources be expected to outdo DeepMind and
OpenAI?
Yet the algorithms that I study run on a student laptop, are trained in a weakly
supervised manner, take seconds to train, and generalise robustly from single
examples without any sort of pre-training whatsoever (for instance, the approach
I study, Meta-Interpetive Learning, comfortably and routinely learns recursive
prorgams with arbitrary structure from a single example, without an example of
the "base-case"). These are capabilities unheard of in neural networks that must
be trained on hundreds of millions of examples [2] of actual programs (rather
than examples of inputs and outputs, as in ILP) in order to generate programs
and cannot solve programming tasks, nor generate solutions, unlike the examples
tasks and solutions in their training sets; while in the case of ILP every
programming task solved is, by definition unseen (hence, "weakly supervised"
because examples are programming tasks without examples of the solutions).
That is what it means for a field to be stalled. When a single graduate
student with a seven-year old laptop can solve problems that the dominant
approach cannot even attetmpt (in this case, learning from a single example, without pretraining and with weak supervision).
I'm not a biochemist. I don't know how much knowledge of how protein fold is important as opposed to knowing the final structure. My understanding is that current expensive method of X-ray crystallography in the range of $100k per protein also does not explain how protein fold, however the structure itself appear to be useful.
Many complex problems do not appear to be explainable in the way planet motions are explainable. The fact that many aspects of the universe are explainable in simple laws is rather extraordinary in itself. In reality, there may be just a lot of irreducible complexity. Maybe success of Deep Learning teaches us that certain things do not fit into framework of science that you assume to be the only correct framework. Reality does not care about what we would want to be true.
Maybe invention of AlphaFold will lead to better understanding of protein folding the same way that invention of steam engine lead to the laws of thermodynamics. Historically, science often lags behind progress of technology. Time will judge.
I also thought that the lack of resources will be your answer. In 2004-2009 neural networks were in quite similar position to ILP. However, NN showed significant potential as evaluated by the broad community at the time and so resources allocated to NN started to grow exponentially. ILP and symbolic methods more broadly do not show similarly convincing potential to the broad community. I'm sure OpenAI, DeepMind, Google, Micosoft etc. would not hesitate to dump equivalent resources they do into many other projects if they saw potential. DeepMind, Microsoft and Google especially hire people with very broad range of expertise.
What may be true is that the world is starting to reach limits of the resources that can be dedicated to deep learning. Exponential growth in resource use can not go forever unless deep learning will start feeding back into available resources and we are certainly not there, not yet.
> Yet the algorithms that I study run on a student laptop, are trained in a weakly supervised manner, take seconds to train, and generalise robustly from single examples without any sort of pre-training whatsoever
>> Maybe invention of AlphaFold will lead to better understanding of protein
folding the same way that invention of steam engine lead to the laws of
thermodynamics. Historically, science often lags behind progress of technology.
Time will judge.
It's not impossible but I'm concerned that neural networks will only be used as
an excuse to not have to understand anything anymore (not just in biology, in
all of the sciences). That would be a terrible outcome- a dumbing down of
science and an eventual loss of our ability to understand the world. As you say,
time will be the judge of that.
I had a look at ARC back when François Chollet's paper came out that presented
it. I was interested because a) it's the kind of problem that ILP eats for
breakfast and b) Chollet's name is well-known and it would be a chance to get
our work noticed by people who normally wouldn't notice it.
However.
ARC is proposed as a benchmark that is hard for current big-data approaches and
that would need elements of intelligent reasoning to solve (if I understand
correctly, that's why you brought it up yourself?). Such benchmarks have been
proposed before, in particular the Bongard problems [1], in machine vision, and the
Winograd schemas [2], in NLP. As with ARC, the "defense" of such datasets against
dumb, no-reasoning, big-data approaches is that there are few examples of each
problem task. And yet, in both Bongard problems and Winograd schemas, neural
nets have now achieved high accuracy. How did they do it? They did it by
cheating: instead of training on the original, and few, Bongard problems, for
example, people created generators for similar problems that could generate many
thousands of problems [3]. That way, the neural nets folks had their big data
and they could train their big networks. Same for Winograd schemas.
... And the same thing has already happened, at a preliminary stage, with ARC:
In the paper I link above, the authors use a data augmentation technique that
consists of rotations and colour transformations of problem tasks. It should be
obvious that this adds no useful information that can help any attempt at
solving ARC problems with reasoning, and that it only serves to help a neural
net better overfit to the training tasks. And yet, the system in the paper
achieves good results in a small selection of ARC tasks. Admittedly, that is a
very small selection (only 10x10 grids and not on any of the held-out test set
that only François Chollet has access to) but the important point is that it is
possible to make progress in solving ARC without actually showing any reasoning
ability, despite the claims of the Kolev et al. paper above; and despite
Chollet's intent for ARC to avoid exactly that.
This is pretty standard in deep learning and the reason why I gripe about
benchmarks in my earlier comment. Deep learning systems consistently beat
benchmarks designed to test abilities that deep learning systems don't have.
They do this by cheating, enabled by weaknesses in said benchmarks. They cheat
by overfitting to surface statistical regularities of the data (see earlier
cited papers) thus learning "shortcuts" around the intended difficulties of the
benchmarks. The trained systems are useless against any other dataset and their
performance degrades precipitously once exposed to real-world data, but that
doesn't matter, because a team that trains such a system to a new SOTA result
will get to publish a paper and claim its academic brownie points, and it's very
difficult to argue with tis "results" without beating the benchmarks yourself,
because of the publishing climate described in Hinton's quote, above.
I mentioned ARC because based on your description it should be "the kind of problem that ILP eats for breakfast". But based on what you say about required effort it does not appear to eat it for breakfast and requires significant engineering fine-tuning, so I call foul on your previous description of your work.
I agree, to a degree, with your assessment of the problem of cheating. Currently community is applying two fixes to that: 1) require few-shot learning ability 2) require good performance on many different few-shot learning benchmarks
My hobby is future forecasting and I was thinking for some time about creating a benchmark for AGI based on Bongard problems, ARC and Montezuma's Revenge. I did have a question about ARC that managed to surprise community: https://www.metaculus.com/questions/3762/will-early-2020-ai-... I have one about Montezuma's Revenge where I require some reasonable score after 30 min of game play, no specific to Montezuma's Revenge engineering allowed, not even retrieving rewards from internal game memory: https://www.metaculus.com/questions/5460/ai-rapidly-learning... I'm aware of the paper you mentioned around ARC and my fix to the benchmark would be to require few-shot generalization on task-level e.g. tested system, if it is a learning system, must have access only to couple of tasks for training before start of evaluation. The same with regard to Bongard problems. Requiring that all 3 benchmarks are solvable by one unified system IMO would be a reasonable test for general, visual intelligence.
DeepMind recently decided to include comparison of Gopher performance on Massive Multitask Language Understanding benchmarks vs. forecasters prediction. Forecasting communities seems to have gained some clout in AI circles (OpenAI is also cooperating with Metaculus on AI forecasting). Few-shot ARC forecast from Metaculus could give you an external estimation of the difficulty that you could use against reviewer 2.
Also, despite what you think about large language models memorizing everything few-shot SuperGlue is still standing: https://www.metaculus.com/questions/4932/when-will-ai-achiev... My intuitive evaluation of GPT-3 light text reasoning skills based on playing around with it matches few-shot SuperGlue score of 72 vs. 90 for avg. human.
> Meanwhile, some big, 30-person team at a large tech corp with a few milllion dollars' budget will be throwing a gigaflop of compute and a few terrabytes of data on the problem and "solving" it without having to learn any "core priors", by finding some way around the dearth of training data, just as was done with Bongard problems and Winograd schemas.
Compute is cheap and it reliably gets cheaper. Models like GPT-3 even with cost of tens of millions of dollars are economically justifiable. It's quite possible that OpenAI is currently turning profit on GPT-3. Real market performance is the ultimate benchmark of whether "performance degrades precipitously once exposed to real-world data".
>> I mentioned ARC because based on your description it should be "the kind of
problem that ILP eats for breakfast". But based on what you say about required
effort it does not appear to eat it for breakfast and requires significant
engineering fine-tuning, so I call foul on your previous description of your
work.
Yep, you caught me. ILP is not magick. You actually have to do
some work to train ILP algorithms.
To clarify, there is no "engineering fine-tuning" required. ILP approaches rely
on background knowledge, which can be thought of as a library of sub-routines
from which target programs are composed. ILP is unique in that the
representation of background knowledge is the same as that of target programs
(they're all logic programs) so they can learn their own background knowledge
incrementally. So the effort to solve ARC would include training an ILP system
to learn core priors from some initial set of low-level primitives. Once
learned, core priors would be included in the background knowledge needed to
actually solve ARC tasks (along with other stuff like rotations, translations
etc that don't have to be learned because they're straightforward to hand-code).
Clearly, that would take time and work. Deep learning approaches simply replace
time and effort with data and compute. But there is always a cost to be paid.
You can't have learning for free.
Learning core priors is not going to be simple. First of all, we don't even know
exactly what they look like. Chollet gives a list, which I think is not
exhaustive, but there's no dataset with examples that can be used in training.
So there's a lot of work to be done, even just to figure out how to learn the
core priors. Or you could do like on Kaggle and solve a sub-set of tasks with
some geometric manipulations and a brute-force search. But that, too, would be
cheating and it wouldn't generalise well anyway.
>> Requiring that all 3 benchmarks are solvable by one unified system IMO would be
a reasonable test for general, visual intelligence.
I don't think that addresses the real issue with benchmarks. Machine learning
benchmarks don't work because we have no idea what we want them to test, and
that in turn because we don't have a theory of intelligence and artificial
intelligence. Benchmarks like the Bongard problems, or ARC, now, are just
educated guesses about the kind of behaviour an intelligent system should
exhibit. Because they are just guesses, we're not really going anywhere and
every few years someone comes up with a new benchmark, claims that this one is
better than earlier ones, and then watches in horror as some dumb machine rips
it to shreds. This will continue for ever until someone comes up with a
plausible theory of what, exactly, it is that we are trying to test. Chollet
actually tried that, but I think wrapping it all up in yet another new benchmark
removes the focus from the theoretical work and focuses it on racing for SOTA
again, thus exposing his work to the same old big-data shennanigans.
For example, 72% accuracy on a language understanding task should indicate
pretty strong language understanding ability already _if_ the benchmark was
actually testing language understanding. 72% accuracy would certainly beat a
dog, or a chimp, it would probably beat a toddler and may even beat an educated
adult who is not a native speaker of the test language. So it's clear that
testing "few" or "any" shot makes no difference because no language
understanding ability is being tested anyway. The benchmarks are fundamentally
broken. They won't get any better by tweaking them. You can't make a thermometer
measure luminosity no matter how much you tweak it.
Anyway pre-training, as I noted earlier, removes all legitimacy from any claims
of "few shot" learning. First we need to undersand what is being learned during
pre-training and if that has anything to do with some element of intelligent,
then we can decide what &...
That statement was in the context of implementing well known algorithms rather than researching unknown ideas. ART is totally a new approach in comparison to backward prop, hope it inspires some to continue research in this area. Deep Learning has so many use cases yet to be implemented that has business value but from the point of research its close to saturation. Time to explore new ideas.
Can't be trusted => can't be fully trusted != is useless
Also nothing is perfect.
And even severely imperfect things can be reasonable good solution/improve a situation.
So don't base a curt ruling on it.
Be very careful about using it when it can put lives at risk (e.g. self driving).
EDIT:And make the people selling the system responsible for it's failure!! You can't expect a person you tell to not touch the wheel for hours to then be able to react asap if something goes wrong, that's not how humans work. Hence putting the responsibility on the human which happens to sit behind the wheel is completely unreasonable.END EDIT
Anyway my main problem with DL is that its hard to evaluate how trust-able it is if applied to a very diverse real world situations (i.e. self driving). You can evaluate it on data sets but data sets are extremely unlikely to get anywhere close to representing a diverse reality, and while each edge case not represented is very rare due to the diversity of reality there is a endless number of them so sooner or later they will be hit. And as they are edge case not represented by your data or evaluation sets you can't relay reason well what will happen if they are it. Furthermore due to the nature of data collection it's not rare for independent data sets to unexpectedly share subtle biases.
A good example is a Tesla car getting confused because in front of it was a truck which had traffic lights loaded on the back. A extremely rare/unusual situation, which happened anyway.
So you have to layer different systems, preferable also some fallback emergency systems not dependent on DL.
I'd like to see some working applications w/ datasets that could be played around with and tweaked, apart from just the papers he's published. (And even then the linked page from the article is largely broken, searching or browsing by author returns an error for me and I'm not paging through 26 pages of listing to find one of his papers)
Without that, this sounds interesting but isn't something I can really evaluate beyond "researcher claims his research is better than status quo"
That's true of traditional models as well. The difference is that when there's a screw up it might be a little easier to determine the cause. That is certain useful, but if there are, on average, fewer screw ups with DL then the net benefit would seem to favor DL.
What I'd really like is to have direct access to a working ART model and dataset for it... I wasn't able to find one after a bit if searching though. (I am however on mobile at the moment which makes that a bit more cumbersome)
This is a legit researcher, but the article is mostly a "why mine is better" without justification. There must be something better to read on this.
The performance problem with deep learning is well known. Great results up to 90%+ of the time, totally bogus a few percent of the time. Great for ad targeting. Automatic driving, not so much. If somebody got past that, it's important.
Deep learning based systems are at least trustable as working animals I think... They share a lot of similarities. You can not know how it thinks, but you can train them, observe them, and then trust them to a certain level.
Disclaimer: I have a PhD from the department Stephen Grossberg founded. Gail Carpenter, the co-creator of ART, was my advisor. I am also co-authoring a neuroscience non-fiction book that features Steve's work, along with other stuff. (Just so you know I am a reasonably biased observer).
Most of you have probably not heard of Stephen Grossberg. He is in my opinion one of the greatest neuroscientists of this generation. Maybe the greatest theoretical neuroscientist ever.
Neuroscience is swimming (barely) in a ton of data, but very little in the way of a unifying framework. Steve's body of work over 60 years provides that. ART is only one part of it. ART has some way to go before proving itself as undeniably useful as deep-learning on the application side of things. But it is very compelling as a model for how our (mammalian) brains might be operating.
One fundamental idea, which is worth considering in applications is how almost all of deep-learning is founded on the idea of an "error" (A mis-match between expectation and the prediction) and how this flows through the system. Backprop is obviously the most popularly known method for this flow. ART is fundamentally different in that it works with the idea of a match rather than a mis-match. This foundational even philosophical difference is interesting because if allows for learning to happen all the time even while perceiving is simultaneously happening. It also allows for learning to be unsupervised. This philosophical departure point is actually very interesting: If you rely entirely on the idea of an error and throw out the rest, can you even hope to build a conscious system?
Steve btw also has what I think is the most elegant and mechanistically precise explanation of consciousness, which builds on ART. Even if you are skeptical of the current promise of ART as a replacement for deep-learning, his explanation of how ART-like systems enable consciousness (and different conscious percepts) are alone worth the price of admission.
I remember ART from 25 years ago, sounded very promising even if I didn't fully understand at the time (nor now, but not revisited).
Obviously back-propagation based NN have become very popular (even got a rebrand as Deep Learning), how has ART changed in the last 25 years, is there any interesting results using ART instead of DL?
There have been updates on the supervised versions of ART (ARTMAP) but none I know of, which have been shown to be as as effective as DL methods on very large datasets. Not sure if there have been many/any attempts to tackle large datasets. DL methods, while opaque, often result in good feature basis sets. ART given it was devised to model perception in brains, assumes features are created by lower network layers (V1, primary auditory cortex etc). I don't think there's any reason to suspect ART could not be extended to aid feature discovery, but it hasn't been done as far as I know.
It's weird how many unrecognized geniuses there are in the neuroscience all claiming to have explained the brain and people just need to recognize their genius. All sounds very similar to Jeff Hawkins.
Your skepticism is warranted :) Grossberg is not that unrecognized. I'd recommend Talking Nets by Anderson and Rosenfeld and Margaret Boden's Mind as Machine. Both are scholarly histories of the field. Boden's book is far more comprehensive. Worth reading in any case if you are interested in the study of the brain and mind.
66 comments
[ 0.30 ms ] story [ 127 ms ] threadI don't know about this ART thing though, never heard of it.
Sure the SCII AI eventually became very strong but it’s also somewhat immersion breaking to see all the odd stuff it ends up doing. And that’s with a well tuned game, in something under active development you end up with a host of other issues.
I mean, playing well is obviously not a goal for making the AI for a single-player mode or a bot for a multi-player game; the goal there is to play in an entertaining way, and the optimal system probably is one that is simple, easy to test, and is beatable in a manner that's fun for the players. Probably the only niche (very narrow!) where you would actually want near-optimal AI is for making a "sparring partner" for training e-sports players.
People working on AI systems that are going to be actually used in games shipped to gamers is a separate community with (IMHO) almost no overlap.
For all we know, there are multiple ways to converge to a trainable input recognition system. And if ART is bringing comprehensibility to the table, it may look even less like what actual brains do than modern deep learning algorithms; nothing about what I know regarding how brains operate suggests to me that they are comprehensible as anything other than "a pattern that works."
All that to say, it's easy to find flaws with it, but it has been game changing. I'm naturally skeptical about someone who has just published a book on something hes been working on for 50 years that he believes is "better" than the current (admittedly fad) technology. I'd like to see some real examples about the specific kinds of problems his tech solves, and how it solves them better than other methods. If he's saying it's better than deep learning on things deep learning is good at (e.g. can it beat a neural network model at classifying imagenet?) then I call BS absent a compelling demo. If it's better at some other thing, let's hear what that is.
I think the progress was the result of competitions that kept everyone blindfolded to test datasets. No one could exaggerate the performance of their methods, allowing deep learning to emerge as a winning solution. I think deep learning has fallen short of the hype because researchers are exaggerating their results, and what is needed to propel progress is a renewed focus (with funding) for competitions.
Advances in DL have been mostly empirical not theoretical from what I have seen. It's less about being able to "envisage" and more about empirical testing.
I may be reading too deeply into your question and you don't actually mean that theoretical. For something lighter, I'd suggest studying some of the common model architectures in the area you're interested in (like you can look at the pytorch lightning or torchvision or detectron2 models or etc source code) and understanding why things are the way they are.
[0] https://topml.rice.edu/ (this is an annual conference, can look at last years talks and references therein, search for topml, etc)
You will be very disappointed if you expect that there is some grand theory of Deep Learning.
My suggestion is to follow Arxiv Sanity: http://www.arxiv-sanity.com/top
We have these tools that solve specific problems in specific contexts yet all these companies are on the hype marketing train. Are these investors not doing due diligence or am I missing something?
Look, everyone will jump on the ART bandwagon if it ever actually beats deep learning. Machine Learning is a very empirical field! But all the people throwing rocks at Deep Learning have failed to propose anything that actually works better.
I think the latest IEEE articles have been a little intellectually lightweight, perhaps we're just simply not their audience. Still, in the style of IEEE papers, I would like to see references for the claims made.
> But all the people throwing rocks at Deep Learning have failed to propose anything that actually works better
I was thinking something similar. DL for sure has problems, but I am yet to see something better. ART doesn't just need to be better, it needs to be better in _at least_ one of: accuracy, training time, recall time, amount of training data, quality of training data.
Extremely high reviews on Amazon: https://www.amazon.com/Conscious-Mind-Resonant-Brain-Makes/p...
Anecdotal example: a few months ago I bought an electric shaver that was very inexpensive and had thousands of 5-star reviews. It failed two months later, and when I went to leave the negative review, I found that the product no longer existed, but another identical-looking product with a different name from a different(?) company was the best-reviewed cheap electric shaver.
Edit: not claiming or implying anything about the author, book or the reviews. Just pointing the curve doesn't look legit to me.
Please don't explain me why your model is better.
_show me_ your model is better with results.
I see a lot of: "this model is explainable and has no forgetting modes". But I cannot care about an explainable model that does not work.
There are many classic explainable models, how do they compare to art in real life applications?
"The problem with today’s AI, he says, is that it tries to imitate the results of brain processing instead of probing the mechanisms that give rise to the results."
Today's AI works really well for many applications (computer vision has leaped in quality). But of course I did not expect perfection from ImageNet trained models. I also did not expect gpt3 to be agi. But are they good (especially when compared with what we had not long ago)? Hell yeah
Also, the advancements in protein folding are superb.
I understand the author is not dismissing the current advancement. But if you claim your model is better you ought to compare it with the state of the art
Hard to assess vague claims without seeing something concrete.
Literally nothing in that repository makes an argument that this method is better than deep learning in any aspect whatsoever, they don't even attempt such a comparison.
Regarding a direct comparison, it may be difficult to do such a comparison by just running some code on data sets, which the one group of algorithms has been heavily optimized for with excessive funding, while the other has been not. Even more so, as such a review would have a difficult time to asses any of the critical features, like chances of memory loss. Comparing technologies isn't as easy as running them side by side.
Let's see the performance of ART on some benchmark tasks, then we can talk.
Here's Geoff Hinton on benchmarks (again; I cite that all the time):
GH: One big challenge the community faces is that if you want to get a paper published in machine learning now it's got to have a table in it, with all these different data sets across the top, and all these different methods along the side, and your method has to look like the best one. If it doesn’t look like that, it’s hard to get published. I don't think that's encouraging people to think about radically new ideas.
Now if you send in a paper that has a radically new idea, there's no chance in hell it will get accepted, because it's going to get some junior reviewer who doesn't understand it. Or it’s going to get a senior reviewer who's trying to review too many papers and doesn't understand it first time round and assumes it must be nonsense. Anything that makes the brain hurt is not going to get accepted. And I think that's really bad.
What we should be going for, particularly in the basic science conferences, is radically new ideas. Because we know a radically new idea in the long run is going to be much more influential than a tiny improvement. That's I think the main downside of the fact that we've got this inversion now, where you've got a few senior guys and a gazillion young guys.
https://www.wired.com/story/googles-ai-guru-computers-think-...
And here's a recent paper noting yet another way machine learning benchmarks are borked:
https://arxiv.org/abs/2003.08907
P.S. And, please leave that aggressive attitude out of research discussions ("bullshit walks"?). Science is not the constant comparison of dick measurements. That's perhaps the point of sports, but anything that stifles debate and inhibits the development of new ideas is anti-science.
Just last year there was a breakthrough in decades old problem of protein-folding. The year before that we figured out that large language models have astonishing few shot learning abilities. CLIP introduced strong 0-shot capabilities to image recognition. Progress in image generation from descriptions has been nothing but astonishing with Dall-E and GLIDE from OpenAI.
Not to mention breakthroughs in game theoretical problems like Counterfactual Regret Minimization and more recent application of Deep Learning to make CFR practical in actual games. Including continuous progress in the AlphaGo linage to AlphaZero, MuZero, and last year EfficientZero and Player of Games.
To make an analogy it's as if we could sort ever larger lists because we keep sorting them with ever bigger computers... but the only sorting algorithm we know is bubblesort.
You bring up large language models. Language models are about as old as computational linguistics [6]. Large language models are... large. There's nothing that's new about them. "Attention" is an anthropomorphically misnamed engineering tweak of the kind that was rightly lambasted by Drew McDermot fourty years ago [7]. Self-play and reinforcement Learning are now more than 60 years old [8].
As to "few shot" learning, that is said for a language model previously trained on a copy of the entire web! In what sense is that "few shot"? It is truly shocking to hear researchers make such obviously false claims with a straight face. And that they are rewarded with publication of their work is nothing less than a scandal.
Deep learning research is going nowhere. It is stuck in a rut. It is dead. It has given up the spirit. It is an ex-research field. This research field is a goner. It is dead, jim. And I'm sure it will be decades before people start to say it out loud (the current generation of reserachers will leave the dirty job to their students), but that doesn't change a thing.
______________
[1] Except in identifying new weaknesses of deep learning systems, as for example Overinterpretation in the paper I cite above, adversarial examples [2], shortcut learning [3], "the elephant in the room" [4], being "right for the wrong reasons" [5] etc etc etc etc.
[2] https://arxiv.org/abs/1312.6199
[3] https://www.nature.com/articles/s42256-020-00257-z
[4] https://arxiv.org/abs/1808.03305
[5] https://aclanthology.org/P19-1334.pdf
Which of course are very welcome. The only way to address the limitations of current approaches is to become aware of them and try to understand them. Unfortunately, it is only a tiny, tiny minority in the field who does that, while the majority is happy to spout nonsense like "[a billion plus] few shot" learning etc.
[6] https://www.semanticscholar.org/paper/Speech-and-language-pr...
[7] https://www.semanticscholar.org/paper/Artificial-intelligenc...
[8] Arhtur Samuel's checkers player beat a human champion in 1961. Donald Michie build MENACE, a reinforcement learning algorithm implemented in matchboxes in 1960:
https://rodneybrooks.com/forai-machine-learning-explained/
Ctrl+F for "Machine learning started with games" to find the relevant section.
But you seem well read, so I don't think there is a point in trying to convince you.
It's also interesting coming from someone researching Inductive Logic Programming that provides no results at all.
From my perspective, I don't care for theoretical advances. I care about results and if scale provides results, I'm happy.
For people who are just reading this exchange take a look here:
https://pbs.twimg.com/media/FHHPjU0VIAUrSMA?format=jpg&name=...
It's 16 computer generated images from the description below each image. Having capability like that is just very useful and fun.
I wish you also luck in finding a "cat paying checkers in style of Salvador Dali" or a "psychedelic painting of a hamster dragon" on the internet.
There is a lot of stuff on the web, but not everything.
Like I say you can sort ever larger lists with ever bigger computers and bubblesort. But you can go a lot further if you use the brain and come up with a better sorting algorithm.
But I think you would be really surprised to find out how much humans have achieved by using their brains rather than ever bigger computers. Here's a small example that happened to be on my mind recently:
https://en.wikipedia.org/wiki/Copernican_Revolution
That is an example of the difference between expensive toys like Dall-E and world-changing scientific work. Though I appreciate that the difference can sometimes appear somewhat mudddled and a machine that can generate a "hamster dragon" may seem as an important scientific breakthrough. Depends on what you're more interested in: hamster dragons, or understanding how the world works. Oh well.
As to this:
>> It's also interesting coming from someone researching Inductive Logic Programming that provides no results at all.
That's an obvious -and odious- attempt at trolling me, but unlike yours, my HN profile is linked to my real-world identity and so I am forced to respond with politenes and to represent my field with dignity. Therefore, I will point out a famous achievement of ILP with which you are sadly unfamiliar and that I expect you to immediately discount on the grounds that you didn't know about it, so it can't have been all that important. I'm linking to an article in the popular press that will be more accessible. It doesn't go into depth over the machine learning approach used but it's ILP; I link to the scholarly article immediately after:
Robot Makes Scientific Discovery All by Itself
For the first time, a robotic system has made a novel scientific discovery with virtually no human intellectual input. Scientists designed “Adam” to carry out the entire scientific process on its own: formulating hypotheses, designing and running experiments, analyzing data, and deciding which experiments to run next.
https://www.wired.com/2009/04/robotscientist/
And the scholarly article:
Functional genomic hypothesis generation and experimentation by a robot scientist
https://www.nature.com/articles/nature02236
There isn't really that much difference. DeepMind took some 3 years to max out CASP protein-folding benchmark running from 1994 on a problem known since ~1960. Protein folding is exactly the type of world-changing scientific work. But it is just a tip of the iceberg. DL is being successfully applied to PDE solving or electron density predictions and more.
The psychedelic hamster dragon is an example of the fact that DL is capable of 0-shot generalization. It is a significant scientific observation in itself.
And with regard to your example. We are discussing whether deep learning has staled. I give you mostly examples from last year like EfficientZero (30 Oct) ~500x improvement in sample efficiency over 2013 DQN, some literally not older than a month: GLIDE (20 Dec), Player of Games (6 Dec).
You are giving me example from 2004 (2009?).
I'm not trying to troll you. I just think you are biased in you assessment of significant results and what it means to be stalled.
Nope. Not at all. That's a misunderstanding of the goal of science which is to explain how the world works. AlphaFold, like all neural network models is a predictive model but has no explanatory power. It can predict the structure of proteins from sequences but it cannot explain why, or how, proteins fold. Scientists can still not explain how proteins fold, certainly not by interacting with AlphaFold.
Why is that a problem? You said you care about results. I linked to the wikipedia article about the Copernican Revolution. I recommend you read the wikipedia article on epicycles [1]. To summarise, for hundreds of years, from Hipparchus of Nicaea, through Claudius Ptolemy and all the way up to Copernicus, astronomers used systems of epicycles (circles-upon-circles) to explain the apparent movements of the planets. Epicycle-based models matched observations very well and predicted the movement of the planets very accuratey because it is always possible to fit a smooth curve with arbitrary precision given a sufficient number of epicycles. However, even Copernicus' model that had the sun in the center of the universe, unlike many earlier models, could not explain why the planets moved the way they did. A true explanation only became available when Isaac Newton developed his laws of universal gravitation.
Newton's theory (itself later superseded by Einstein's theory of general relativity) provided a true explanation of observable phenomena and a way to calculate better models and obtain more accurate predictions than were ever possible before. That is the power of explaining and understanding the world and simply modelling an incomprehensible set of observations, which is all that neural networks can do, doesn't even come close.
>> I'm not trying to troll you. I just think you are biased in you assessment of significant results and what it means to be stalled.
ILP is a tiny field with maybe a few dozen people that consistently publish on the subject. By comparison, there are many tens of thousands of papers published on deep learning every year, the deep learning conferences receive many thousands of submissions and research is funded with many millions of dollars by governments, militaries, and the largest of large technology corporatations, the Googles and Facebooks etc, and training state-of-the-art models requires many terrabytes of data, many petaflops of computing power, to the extent that training such models is now only feasible for the aforementioned large corporations. Normally, I'd complain that you are comparing apples to oranges. How can a few people with meager resources be expected to outdo DeepMind and OpenAI?
Yet the algorithms that I study run on a student laptop, are trained in a weakly supervised manner, take seconds to train, and generalise robustly from single examples without any sort of pre-training whatsoever (for instance, the approach I study, Meta-Interpetive Learning, comfortably and routinely learns recursive prorgams with arbitrary structure from a single example, without an example of the "base-case"). These are capabilities unheard of in neural networks that must be trained on hundreds of millions of examples [2] of actual programs (rather than examples of inputs and outputs, as in ILP) in order to generate programs and cannot solve programming tasks, nor generate solutions, unlike the examples tasks and solutions in their training sets; while in the case of ILP every programming task solved is, by definition unseen (hence, "weakly supervised" because examples are programming tasks without examples of the solutions).
That is what it means for a field to be stalled. When a single graduate student with a seven-year old laptop can solve problems that the dominant approach cannot even attetmpt (in this case, learning from a single example, without pretraining and with weak supervision).
_______...
Many complex problems do not appear to be explainable in the way planet motions are explainable. The fact that many aspects of the universe are explainable in simple laws is rather extraordinary in itself. In reality, there may be just a lot of irreducible complexity. Maybe success of Deep Learning teaches us that certain things do not fit into framework of science that you assume to be the only correct framework. Reality does not care about what we would want to be true.
Maybe invention of AlphaFold will lead to better understanding of protein folding the same way that invention of steam engine lead to the laws of thermodynamics. Historically, science often lags behind progress of technology. Time will judge.
I also thought that the lack of resources will be your answer. In 2004-2009 neural networks were in quite similar position to ILP. However, NN showed significant potential as evaluated by the broad community at the time and so resources allocated to NN started to grow exponentially. ILP and symbolic methods more broadly do not show similarly convincing potential to the broad community. I'm sure OpenAI, DeepMind, Google, Micosoft etc. would not hesitate to dump equivalent resources they do into many other projects if they saw potential. DeepMind, Microsoft and Google especially hire people with very broad range of expertise.
What may be true is that the world is starting to reach limits of the resources that can be dedicated to deep learning. Exponential growth in resource use can not go forever unless deep learning will start feeding back into available resources and we are certainly not there, not yet.
> Yet the algorithms that I study run on a student laptop, are trained in a weakly supervised manner, take seconds to train, and generalise robustly from single examples without any sort of pre-training whatsoever
Show me that applied to: https://github.com/fchollet/ARC
It's not impossible but I'm concerned that neural networks will only be used as an excuse to not have to understand anything anymore (not just in biology, in all of the sciences). That would be a terrible outcome- a dumbing down of science and an eventual loss of our ability to understand the world. As you say, time will be the judge of that.
>> Show me that applied to: https://github.com/fchollet/ARC
I had a look at ARC back when François Chollet's paper came out that presented it. I was interested because a) it's the kind of problem that ILP eats for breakfast and b) Chollet's name is well-known and it would be a chance to get our work noticed by people who normally wouldn't notice it.
However.
ARC is proposed as a benchmark that is hard for current big-data approaches and that would need elements of intelligent reasoning to solve (if I understand correctly, that's why you brought it up yourself?). Such benchmarks have been proposed before, in particular the Bongard problems [1], in machine vision, and the Winograd schemas [2], in NLP. As with ARC, the "defense" of such datasets against dumb, no-reasoning, big-data approaches is that there are few examples of each problem task. And yet, in both Bongard problems and Winograd schemas, neural nets have now achieved high accuracy. How did they do it? They did it by cheating: instead of training on the original, and few, Bongard problems, for example, people created generators for similar problems that could generate many thousands of problems [3]. That way, the neural nets folks had their big data and they could train their big networks. Same for Winograd schemas.
... And the same thing has already happened, at a preliminary stage, with ARC:
https://arxiv.org/abs/2011.09860
In the paper I link above, the authors use a data augmentation technique that consists of rotations and colour transformations of problem tasks. It should be obvious that this adds no useful information that can help any attempt at solving ARC problems with reasoning, and that it only serves to help a neural net better overfit to the training tasks. And yet, the system in the paper achieves good results in a small selection of ARC tasks. Admittedly, that is a very small selection (only 10x10 grids and not on any of the held-out test set that only François Chollet has access to) but the important point is that it is possible to make progress in solving ARC without actually showing any reasoning ability, despite the claims of the Kolev et al. paper above; and despite Chollet's intent for ARC to avoid exactly that.
This is pretty standard in deep learning and the reason why I gripe about benchmarks in my earlier comment. Deep learning systems consistently beat benchmarks designed to test abilities that deep learning systems don't have. They do this by cheating, enabled by weaknesses in said benchmarks. They cheat by overfitting to surface statistical regularities of the data (see earlier cited papers) thus learning "shortcuts" around the intended difficulties of the benchmarks. The trained systems are useless against any other dataset and their performance degrades precipitously once exposed to real-world data, but that doesn't matter, because a team that trains such a system to a new SOTA result will get to publish a paper and claim its academic brownie points, and it's very difficult to argue with tis "results" without beating the benchmarks yourself, because of the publishing climate described in Hinton's quote, above.
Of course, if you w...
I agree, to a degree, with your assessment of the problem of cheating. Currently community is applying two fixes to that: 1) require few-shot learning ability 2) require good performance on many different few-shot learning benchmarks
My hobby is future forecasting and I was thinking for some time about creating a benchmark for AGI based on Bongard problems, ARC and Montezuma's Revenge. I did have a question about ARC that managed to surprise community: https://www.metaculus.com/questions/3762/will-early-2020-ai-... I have one about Montezuma's Revenge where I require some reasonable score after 30 min of game play, no specific to Montezuma's Revenge engineering allowed, not even retrieving rewards from internal game memory: https://www.metaculus.com/questions/5460/ai-rapidly-learning... I'm aware of the paper you mentioned around ARC and my fix to the benchmark would be to require few-shot generalization on task-level e.g. tested system, if it is a learning system, must have access only to couple of tasks for training before start of evaluation. The same with regard to Bongard problems. Requiring that all 3 benchmarks are solvable by one unified system IMO would be a reasonable test for general, visual intelligence.
DeepMind recently decided to include comparison of Gopher performance on Massive Multitask Language Understanding benchmarks vs. forecasters prediction. Forecasting communities seems to have gained some clout in AI circles (OpenAI is also cooperating with Metaculus on AI forecasting). Few-shot ARC forecast from Metaculus could give you an external estimation of the difficulty that you could use against reviewer 2.
Also, despite what you think about large language models memorizing everything few-shot SuperGlue is still standing: https://www.metaculus.com/questions/4932/when-will-ai-achiev... My intuitive evaluation of GPT-3 light text reasoning skills based on playing around with it matches few-shot SuperGlue score of 72 vs. 90 for avg. human.
> Meanwhile, some big, 30-person team at a large tech corp with a few milllion dollars' budget will be throwing a gigaflop of compute and a few terrabytes of data on the problem and "solving" it without having to learn any "core priors", by finding some way around the dearth of training data, just as was done with Bongard problems and Winograd schemas.
Compute is cheap and it reliably gets cheaper. Models like GPT-3 even with cost of tens of millions of dollars are economically justifiable. It's quite possible that OpenAI is currently turning profit on GPT-3. Real market performance is the ultimate benchmark of whether "performance degrades precipitously once exposed to real-world data".
Yep, you caught me. ILP is not magick. You actually have to do some work to train ILP algorithms.
To clarify, there is no "engineering fine-tuning" required. ILP approaches rely on background knowledge, which can be thought of as a library of sub-routines from which target programs are composed. ILP is unique in that the representation of background knowledge is the same as that of target programs (they're all logic programs) so they can learn their own background knowledge incrementally. So the effort to solve ARC would include training an ILP system to learn core priors from some initial set of low-level primitives. Once learned, core priors would be included in the background knowledge needed to actually solve ARC tasks (along with other stuff like rotations, translations etc that don't have to be learned because they're straightforward to hand-code). Clearly, that would take time and work. Deep learning approaches simply replace time and effort with data and compute. But there is always a cost to be paid. You can't have learning for free.
Learning core priors is not going to be simple. First of all, we don't even know exactly what they look like. Chollet gives a list, which I think is not exhaustive, but there's no dataset with examples that can be used in training. So there's a lot of work to be done, even just to figure out how to learn the core priors. Or you could do like on Kaggle and solve a sub-set of tasks with some geometric manipulations and a brute-force search. But that, too, would be cheating and it wouldn't generalise well anyway.
>> Requiring that all 3 benchmarks are solvable by one unified system IMO would be a reasonable test for general, visual intelligence.
I don't think that addresses the real issue with benchmarks. Machine learning benchmarks don't work because we have no idea what we want them to test, and that in turn because we don't have a theory of intelligence and artificial intelligence. Benchmarks like the Bongard problems, or ARC, now, are just educated guesses about the kind of behaviour an intelligent system should exhibit. Because they are just guesses, we're not really going anywhere and every few years someone comes up with a new benchmark, claims that this one is better than earlier ones, and then watches in horror as some dumb machine rips it to shreds. This will continue for ever until someone comes up with a plausible theory of what, exactly, it is that we are trying to test. Chollet actually tried that, but I think wrapping it all up in yet another new benchmark removes the focus from the theoretical work and focuses it on racing for SOTA again, thus exposing his work to the same old big-data shennanigans.
For example, 72% accuracy on a language understanding task should indicate pretty strong language understanding ability already _if_ the benchmark was actually testing language understanding. 72% accuracy would certainly beat a dog, or a chimp, it would probably beat a toddler and may even beat an educated adult who is not a native speaker of the test language. So it's clear that testing "few" or "any" shot makes no difference because no language understanding ability is being tested anyway. The benchmarks are fundamentally broken. They won't get any better by tweaking them. You can't make a thermometer measure luminosity no matter how much you tweak it.
Anyway pre-training, as I noted earlier, removes all legitimacy from any claims of "few shot" learning. First we need to undersand what is being learned during pre-training and if that has anything to do with some element of intelligent, then we can decide what &...
Also nothing is perfect.
And even severely imperfect things can be reasonable good solution/improve a situation.
So don't base a curt ruling on it.
Be very careful about using it when it can put lives at risk (e.g. self driving).
EDIT:And make the people selling the system responsible for it's failure!! You can't expect a person you tell to not touch the wheel for hours to then be able to react asap if something goes wrong, that's not how humans work. Hence putting the responsibility on the human which happens to sit behind the wheel is completely unreasonable.END EDIT
Anyway my main problem with DL is that its hard to evaluate how trust-able it is if applied to a very diverse real world situations (i.e. self driving). You can evaluate it on data sets but data sets are extremely unlikely to get anywhere close to representing a diverse reality, and while each edge case not represented is very rare due to the diversity of reality there is a endless number of them so sooner or later they will be hit. And as they are edge case not represented by your data or evaluation sets you can't relay reason well what will happen if they are it. Furthermore due to the nature of data collection it's not rare for independent data sets to unexpectedly share subtle biases.
A good example is a Tesla car getting confused because in front of it was a truck which had traffic lights loaded on the back. A extremely rare/unusual situation, which happened anyway.
So you have to layer different systems, preferable also some fallback emergency systems not dependent on DL.
Without that, this sounds interesting but isn't something I can really evaluate beyond "researcher claims his research is better than status quo"
That's true of traditional models as well. The difference is that when there's a screw up it might be a little easier to determine the cause. That is certain useful, but if there are, on average, fewer screw ups with DL then the net benefit would seem to favor DL.
What I'd really like is to have direct access to a working ART model and dataset for it... I wasn't able to find one after a bit if searching though. (I am however on mobile at the moment which makes that a bit more cumbersome)
The performance problem with deep learning is well known. Great results up to 90%+ of the time, totally bogus a few percent of the time. Great for ad targeting. Automatic driving, not so much. If somebody got past that, it's important.
The Wikipedia article is light on details [1], the 'types' described look basic and preliminary.
[1] https://en.wikipedia.org/wiki/Adaptive_resonance_theory
Most of you have probably not heard of Stephen Grossberg. He is in my opinion one of the greatest neuroscientists of this generation. Maybe the greatest theoretical neuroscientist ever.
Neuroscience is swimming (barely) in a ton of data, but very little in the way of a unifying framework. Steve's body of work over 60 years provides that. ART is only one part of it. ART has some way to go before proving itself as undeniably useful as deep-learning on the application side of things. But it is very compelling as a model for how our (mammalian) brains might be operating.
One fundamental idea, which is worth considering in applications is how almost all of deep-learning is founded on the idea of an "error" (A mis-match between expectation and the prediction) and how this flows through the system. Backprop is obviously the most popularly known method for this flow. ART is fundamentally different in that it works with the idea of a match rather than a mis-match. This foundational even philosophical difference is interesting because if allows for learning to happen all the time even while perceiving is simultaneously happening. It also allows for learning to be unsupervised. This philosophical departure point is actually very interesting: If you rely entirely on the idea of an error and throw out the rest, can you even hope to build a conscious system? Steve btw also has what I think is the most elegant and mechanistically precise explanation of consciousness, which builds on ART. Even if you are skeptical of the current promise of ART as a replacement for deep-learning, his explanation of how ART-like systems enable consciousness (and different conscious percepts) are alone worth the price of admission.
I’ve written more about Stephen Grossberg and his book here: https://saigaddam.medium.com/the-greatest-neuroscientist-you...
Obviously back-propagation based NN have become very popular (even got a rebrand as Deep Learning), how has ART changed in the last 25 years, is there any interesting results using ART instead of DL?